A study led by Biyu He of New York University’s Grossman School of Medicine, with first author Mugihiko Kato, published in *iScience*, found that modern image recognition programs perform significantly worse than human observers at detecting overarching shapes and silhouettes. The researchers compared test subjects with more than 200 deep neural networks of various architectures and training methods, systematically manipulating 240 images from 48 everyday categories, including solid black silhouettes without internal structure and isolated patterns such as fur textures without recognizable contours. None of the more than 200 AI models fully replicated the human recognition pattern across all conditions, and whenever identification depended solely on the overall silhouette, the algorithms consistently performed worse. In a follow-up experiment, outlines filled with many small crosses were still easily identified by humans, but most AI models failed completely, interpreting cat or butterfly silhouettes as “crossword puzzles,” “window grilles,” “chain mail,” or “nematodes.” He emphasized that today’s image recognition models are not as human-like as often assumed, and while systems trained on images and accompanying text achieved more human-like accuracy rates overall, they too lost their advantage once the overall composition was disrupted.